METHOD FOR TRAINING QUANTUM NEURAL NETWORK AND METHOD FOR CLASSIFYING DATA

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001]The present application is a National Stage of International Application No. PCT/CN2024/082598 filed on Mar. 20, 2024, which claims priority to Chinese patent application No. 202310626399.3 entitled “METHOD FOR TRAINING QUANTUM NEURAL NETWORK AND METHOD FOR CLASSIFYING DATA” filed on May 30, 2023, both of which are incorporated herein by reference in their entireties.

TECHNICAL FIELD

[0002]The present disclosure relates to the technical field of quantum computing, particularly to a method for training a quantum neural network and a method for classifying data.

BACKGROUND

[0003]A traditional neural network, also simply referred to as a neural network, is an artificial intelligence system composed of numerous neurons connected by adjustable connection weights, processing capabilities of large-scale parallel processing, distributed information storage, and excellent self-organizing and self-learning.

[0004]With the development of quantum computing science, the neural network has been combined with the quantum computing science to develop a novel quantum neural network. Classical quantum neural network relies on technologies such as quantum state preparation of data and the quantum state data storage. However, the quantum state preparation process often consumes additional time and space complexity, while the quantum state data storage presents a significant technical challenge, resulting in high training difficulty and complexity for the quantum neural network. Furthermore, the classical quantum neural network often requires a large number of qubits to represent different data, and too many qubits not only lead to high computational costs but also cause the barren plateaus phenomenon during training.

SUMMARY

[0005]Embodiments of the present disclosure provide a method for training a quantum neural network and a method for classifying data, which can reduce the training difficulty and complexity of the quantum neural network, reduce the computational cost of the quantum neural network, and reduce the probability of occurring of the barren plateaus phenomenon.

[0034]In a fifth aspect, some embodiments of the present disclosure provide an electronic device. The electronic device includes a processor and a memory storing computer program instructions. The processor, when executing the computer program instructions, implements the steps of the method for training the quantum neural network provided in any one embodiment in the first aspect, or the steps of the method for classifying data provided in any one embodiment in the second aspect.

[0035]In a sixth aspect, some embodiments of the present disclosure provide a computer-readable storage medium storing computer program instructions. The computer program instructions, when executed by a processor, implement the steps of the method for training the quantum neural network provided in any one embodiment in the first aspect, or the steps of the method for classifying data provided in any one embodiment in the second aspect.

[0036]In a seventh aspect, some embodiments of the present disclosure provide a computer program product. Instructions in the computer program product, when executed by a processor of an electronic device, cause the electronic device to perform the steps of the method for training the quantum neural network provided in any one embodiment in the first aspect, or the steps of the method for classifying data provided in any one embodiment in the second aspect.

BRIEF DESCRIPTION OF THE DRAWINGS

[0037]To more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings used in the embodiments of the present disclosure will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038]FIG. 1 is a flowchart of a method for training a quantum neural network provided by an embodiment of the present disclosure;

[0039]FIG. 2 is a network structural diagram of the quantum neural network provided by the present disclosure;

[0040]FIG. 3 is an equivalent diagram of a quantum circuit provided by the present disclosure;

[0041]FIG. 4 is a flowchart of a method for classifying data provided by an embodiment of the present disclosure;

[0042]FIG. 5 is a schematic structural diagram of an apparatus for training the quantum neural network provided by an embodiment of the present disclosure;

[0043]FIG. 6 is a schematic structural diagram of an apparatus for classifying data provided by an embodiment of the present disclosure; and

[0044]FIG. 7 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure.

DETAILED DESCRIPTION

[0045]The features and exemplary embodiments of various aspects of the present disclosure will be described in detail below. To make the purposes, technical solutions, and advantages of the present disclosure clearer and more understandable, the present disclosure will be further described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific embodiments described herein are intended only to explain the present disclosure and not to limit the present disclosure. For those skilled in the art, the present disclosure can be implemented without some of the details described herein. The following description of the embodiments is intended solely to provide a better understanding of the present disclosure by illustrating examples of the present disclosure.

[0046]It should be noted that, in the present disclosure, relational terms such as “first” and “second” are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or sequence between the entities or operations. Moreover, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements includes not only those elements but also includes other elements not explicitly listed or elements inherent to such process, method, article, or device. Without further limitation, the elements defined by the phrase “including . . . ” do not exclude additional identical elements in the process, method, article, or device that includes these elements.

[0047]To address the technical problems of the prior art, the embodiments of the present disclosure provide a method for training a quantum neural network and a method for classifying data. The method for training the quantum neural network can be applied to scenarios for training the quantum neural network. The method for training the quantum neural network provided by some embodiments of the present disclosure is firstly described below.

[0048]FIG. 1 is a flowchart of a method for training a quantum neural network provided by an embodiment of the present disclosure. As shown in FIG. 1, the method for training the quantum neural network specifically includes S110, S120, S130, S140, S150, and S160.

[0049]In S110: sample data for training the quantum neural network and a sample category label corresponding to the sample data are acquired, the quantum neural network including a feature extraction layer, a unitary matrix layer, and a quantum circuit.

[0050]In S120: feature extraction is performed on the sample data using the feature extraction layer to obtain a sample feature.

[0051]In S130: the sample feature is input into the unitary matrix layer to obtain a unitary matrix corresponding to the sample feature.

[0052]In S140: a quantum state of a first qubit is adjusted based on the unitary matrix to obtain a second qubit, the quantum state of the first qubit corresponding to the sample category label.

[0053]In S150: a quantum state fidelity between the second qubit and the first qubit is determined using the quantum circuit, and a loss value is determined based on the quantum state fidelity.

[0054]In S160: a network parameter in the quantum neural network is adjusted based on the loss value, and the operation then returns to execute the acquiring of the sample data for training the quantumneural network and the sample category label corresponding to the sample data until the quantum neural network converges, so as to obtain the trained quantum neural network.

[0055]Therefore, the structure of the quantum neural network is improved by providing the feature extraction layer, the unitary matrix layer, and the quantum circuit in the quantum neural network. Then, during the training process, the sample feature extracted from the feature extraction layer and corresponding to the sample data is directly input into the unitary matrix layer, the unitary matrix corresponding to the sample feature is prepared to replace the preparation of quantum state data in the traditional method, the first qubit serves as a substitute for the activation function, so that the quantum state of the first qubit is adjusted to be flipped by the unitary matrix; after obtaining the second qubit, a similarity comparison between the second qubit and the first qubit is performed by the quantum circuit to calculate the loss function. In this way, in the embodiments of the present disclosure, the quantum neural network can be obtained through training without quantum state preparation, quantum state data storage, or excessive qubits, which can reduce the training difficulty and complexity of the quantum neural network, reduce the computational cost of the quantum neural network, and reduce the probability of occurring of the barren plateaus phenomenon.

[0056]The specific implementation methods for the aforementioned respective steps are described below.

[0057]In some embodiments, in S110, a training set D_train including M pieces of sample data can be provided, where M is an integer greater than 1. On such basis, a feature item in the training set D_train can be represented as {right arrow over (x)}=({right arrow over (x₁)}, {right arrow over (x₂)}, . . . , {right arrow over (x_M)}), where the feature item corresponding to the m-th piece of sample data is {right arrow over (x_m)}, where 1≤m≤M. One piece of sample data can be provided with one sample category label, so that a label item corresponding to the training set D_train can be {right arrow over (y)}=(y₁, y₂, . . . , y_M).

[0058]Additionally, the sample data can be selected according to the usage scenario for the quantum neural network. For example, under a condition that the quantum neural network is configured to predict merchant lending risks, the sample data can be selected from relevant data of risky merchants and relevant data of non-risky merchants, and the sample category labels can include risky and non-risky merchants.

[0059]Exemplarily, after obtaining the training set D_train, each training iteration can retrieve one piece of sample data {right arrow over (x_m)} and one sample category label y_m corresponding to the sample data {right arrow over (x_m)} from the training set D_train.

[0060]In some embodiments, in S120, different from the structure of the traditional quantum neural network, the quantum neural network provided in the embodiments of the present disclosure can include the feature extraction layer, the unitary matrix layer, and the quantum circuit. The feature extraction layer can be a structure in the traditional neural network configured to process conventional data, and can include an input layer and a hidden layer. Specifically, the input layer can be configured to receive data to be input into the quantum neural network, and the hidden layer can be configured to map the input data to another dimensional space to extract a feature of the input data.

[0061]Exemplarily, the sample data can be input into the quantum neural network via the input layer, and feature extraction is performed on the input sample data by the hidden layer, so as to output the sample feature corresponding to the sample data.

[0062]On such basis, under a condition that the sample data includes N feature data points corresponding to N dimensions, in order to further improve the accuracy of feature extraction, in some embodiments, K hidden units can be provided in the feature extraction layer, where N and K are integers greater than 1.

[0063]Exemplarily, each piece of sample data can include feature data points of N dimensions. That is, the m-th piece of sample data in the training set D_train can be represented as {right arrow over (x_m)}=(x_m1, x_m2, . . . , x_mN), and one feature data on the n-th dimension in the m-th piece of sample data is x_mn, where 1≤n≤N. Additionally, the number of the hidden units, K, can be set by the user based on the actual usage scenario for the quantum neural network.

[0064]On such basis, in some embodiments, the above S120 can specifically include: inputting the N feature data points into each of the K hidden units, and performing feature extraction on the N feature data points using each hidden unit, so as to obtain K sample sub-features, the sample feature including the K sample sub-features.

[0065]Exemplarily, a structure of the quantum neural network provided by some embodiments of the present disclosure can be as shown in FIG. 2. For the feature data points 21 of N dimensions in the sample data {right arrow over (x_m)}, the feature data points 21 of N dimensions can be input into each hidden unit in the hidden layer 22.

[0066]For example, the feature data points 21 of N dimensions are input into a hidden unit h₁ in the hidden layer 22, and feature extraction is performed on the N feature data points using the hidden unit h₁ to obtain one sample sub-feature output by the hidden unit h₁; the feature data points 21 of N dimensions are input into a hidden unit h₂ in the hidden layer 22, and feature extraction is performed on the N feature data points using the hidden unit h₂ to obtain one sample sub-feature output by the hidden unit h₂; . . . ; the feature data points 21 of N dimensions are input into a hidden unit h_K in the hidden layer 22, and feature extraction is performed on the N feature data points using the hidden unit h_K to obtain one sample sub-feature output by the hidden unit h_K. In this way, K sample sub-features can be obtained.

[0067]Here, assuming the input is the m-th piece of sample data, a calculation formula corresponding to the k-th hidden unit in the hidden layer can be expressed as the following formula (1):

[0069]In some embodiments, in S130 and S140, an evolution of the quantum state can be described by unitary transformation, and the unitary matrix layer can be configured to perform unitary transformation on the first qubit based on the sample feature.

[0071]Exemplarily, the sample feature can be directly input into the unitary matrix to cause the unitary matrix to be changed to obtain a unitary matrix corresponding to this sample feature, the unitary matrix is then applied to the first qubit to cause the quantum state of the first qubit to be flipped to obtain the second qubit. The first qubit can be a unit qubit and has a quantum state corresponding to a sample category label for the sample data.

$\begin{array}{cc}|\varphi \rangle =\{\begin{array}{cc}|0\rangle ,& y_m\mathrm{belonging}\mathrm{to}A\\ |1\rangle ,& y_m\mathrm{belonging}\mathrm{to}B\end{array}& \left(2\right)\end{array}$

$|0\rangle =\left(\begin{array}{c}1\\ 0\end{array}\right),\mathrm{and}|1\rangle =\left(\begin{array}{c}0\\ 1\end{array}\right).$

[0074]On such basis, under a condition that the hidden layer includes K hidden units, i.e., under a condition that the sample feature includes K sample sub-features output from the K hidden units, in some embodiments, the above S130 can specifically include: inputting the K sample sub-features into the unitary matrix layer to obtain K unitary matrices corresponding to the K sample sub-features.

[0075]Here, one sample sub-feature can correspond and be computed to obtain one unitary matrix.

[0076]On such basis, in some embodiments, an expression for the unitary matrix layer in some embodiments of the present disclosure can be represented by a formula (3):

[0078]Additionally, in some embodiments, a value of each adjustment parameter can be determined based on the data feature for the sample data.

[0079]Exemplarily, α and β each serve as a leading term to prevent data from being too small and are typically set to 1. For example, in merchant transaction data points, since both transaction amounts and percentages are substantial, there is no need to add amplitude α and β. Conversely, data sometimes is small (e.g., input sample feature being 0.1%), which may cause the first qubit to be deflected insignificantly. On such basis, in combination with measurement errors and noise, measurement may be not be performed. In this case, β is set to be 1000 to increase the value, so that the deflection of the first qubit is more obvious.

[0080]Additionally, in some implementations, under a condition that the K unitary matrices is obtained, S140 can specifically include: multiplying the K unitary matrices by the first qubit to obtain the second qubit.

$\left({\prod }_{i=1}^{K}U\left(h_i\right)\right)$

[0092]Here, as shown in FIG. 2, the quantum neural network provided by some embodiments of the present disclosure can include a quantum circuit 24, and the quantum circuit 24 can specifically include a first Hadamard gate 241, a second Hadamard gate 242, and a swap gate 243. The quantum gates can be arranged in the sequence of circuit elements shown in FIG. 2. The first Hadamard gate 241 and second Hadamard gate 242 can be Hadamard gates whose expression is

$H=\frac{1}{\sqrt{2}}\left[\begin{array}{cc}1& 1\\ 1& -1\end{array}\right].$

The swap gate 243 is

$\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\\ 0& 1& 0& 0\\ 0& 0& 0& 1\end{array}\right].$

$H|0\rangle =\frac{1}{\sqrt{2}}(|0\rangle +|1\rangle ),$

$\begin{array}{cc}|{\psi }_1\rangle =|0\rangle |\phi \rangle |\varphi \rangle & \left(5\right)\end{array}$

$\begin{array}{cc}|{\psi }_4\rangle =H|{\psi }_3\rangle =\frac{1}{\sqrt{2}}H(|0\rangle |\phi \rangle |\varphi \rangle +|1\rangle |\varphi \rangle |\phi \rangle )=\frac{1}{2}[|0\rangle (|\varphi \rangle |\phi \rangle +& \left(8\right)\end{array}$$|\phi \rangle |\varphi \rangle )+|1\rangle (|\varphi \rangle |\phi \rangle -|\phi \rangle |\varphi \rangle )]$

$\begin{array}{cc}{\mathrm{Prob}\left(0\right)=❘\frac{1}{2}[|0\rangle \otimes (|\varphi \rangle |\phi \rangle +|\phi \rangle |\varphi \rangle )]❘}^2=\frac{1}{2}\left(1+{❘\langle \varphi |\phi \rangle ❘}^2\right)& \left(10\right)\end{array}$

$\begin{array}{cc}{❘\langle \varphi |\phi \rangle ❘}^2=2\mathrm{Prob}\left(0\right)-1& \left(11\right)\end{array}$

$\mathrm{Prob}\left(0\right)=\frac{1}{2}\left(1+{❘\langle \varphi |\phi \rangle ❘}^2\right)=0.6,$

[0110]In some embodiments, in S160, after determining the quantum state fidelity between the second qubit and the first qubit, the loss value can be calculated according to the following formula (12):

[0113]Additionally, in the embodiments of the present disclosure, a stochastic gradient descent (SGD) method can be employed to adjust the network parameter, which is specifically presented as follows:

$b_{\mathrm{mk}}\leftarrow b_{\mathrm{mk}}-\eta \frac{\partial c}{\partial b_{\mathrm{mk}}}$

[0115]The principle of SGD lies in updating the parameter using only one sample in each iteration update, i.e., updating with only one piece of sample data. In this way, since only one piece of sample data is adopted in each iteration, the training is faster. Compared to non-stochastic algorithms, SGD can utilize information more efficiently, particularly when information is redundant.

[0116]On such basis, in some specific examples, the process for training the quantum neural network can include: initializing a weight W and a bias b, which can be set to 0 or any random number; performing a forward propagation in which

$\frac{\partial C}{\partial b_{\mathrm{mk}}}\mathrm{and}\frac{\partial C}{\partial w_{\mathrm{nk}}}$

are computed with the given piece of sample data {right arrow over (x_m)} in combination with a weight W and a bias b of the hidden layer; performing a backward propagation in which a partial derivative is calculated and the weight W and the bias b are updated; inferring the loss value with a measured value and computing the loss value C; inputting the next piece of sample data {right arrow over (x_m+1)}, repeating the above steps until all pieces of sample data in the training set D_train are processed, and completing one training round. Multiple training rounds are repeated while observing the change of the loss value, and the training stops until the loss value converges.

[0117]A method for classifying data provided by some embodiments of the present disclosure is described below.

[0118]FIG. 4 is a flowchart of a method for classifying data provided by an embodiment of the present disclosure. The trained quantum neural network used in the method for classifying data is obtained according to the method of training the quantum neural network provided in the aforementioned embodiments.

[0119]As shown in FIG. 4, the method for classifying data specifically includes the following steps:

[0120]In S410: target data to be classified is acquired, and inputting the target data into a quantum neural network, the quantum neural network including a feature extraction layer, a unitary matrix layer, and a quantum circuit.

[0121]In S420: feature extraction is performed on the target data using the feature extraction layer to obtain a target data feature.

[0122]In S430: the target data feature is input into the unitary matrix layer to obtain a unitary matrix corresponding to the target data feature.

[0123]In S440: a quantum state of a fifth qubit is adjusted based on the unitary matrix to obtain a sixth qubit.

[0124]In S450: a quantum state fidelity between the sixth qubit and the fifth qubit is determined using the quantum circuit.

[0125]In S460: a category to which the target data belongs is determined based on the quantum state fidelity.

[0126]Thus, the structure of the quantum neural network is improved by providing the feature extraction layer, the unitary matrix layer, and the quantum circuit in the quantum neural network. Then, during the process for data classification using the quantum neural network, the target data feature extracted from the feature extraction layer and corresponding to the target data to be classified is directly input into the unitary matrix layer, the unitary matrix corresponding to the target data feature is prepared to replace the preparation of quantum state data in the traditional method, the fifth qubit serves as a substitute for the activation function to allow the quantum state of the fifth qubit to be adjusted to be flipped by the unitary matrix, after obtaining the sixth qubit, a similarity comparison between the sixth qubit and the fifth qubit is performed by the quantum circuit to obtain the quantum state fidelity, and the category to which the target data belongs id determined based on the quantum state fidelity. In this way, in the embodiments of the present disclosure, data classification can be performed using the quantum neural network without quantum state preparation, quantum state data storage, or excessive qubits, which can reduce the training difficulty and complexity of the quantumneural network, and reduce the computational cost of the quantum neural network.

[0127]The following describes the specific implementation methods for each of the aforementioned steps.

[0128]In some embodiments, the aforementioned method for classifying data can be specifically applied to a scenario for predicting merchant lending risks. In this scenario, the target data to be classified can be data related to the merchant to be predicted.

[0129]Additionally, it should be noted that since steps S420 to S450 are identical or similar to the previously described steps S120 to S150, respectively, and are primarily to replace the sample data in S120 to S150 with the target data to be classified. Therefore, for the sake of brevity, the steps from S420 to S450 will not be elaborated herein, and the explanation provided for steps S120 to S150 also applicable to steps S420 to S450 in this embodiment.

[0131]Furthermore, with respect to the aforementioned S460, in some embodiments, after determining and obtaining the quantum state fidelity between the sixth qubit and the fifth qubit, it is determined whether the target data belongs to the same category as the category label corresponding to the fifth qubit according to a magnitude of this quantum state fidelity. Subsequently, and then the category label to which the target data belongs can be determined by referencing the category label corresponding to the fifth qubit.

$\begin{array}{cc}{❘\langle \varphi |\phi \rangle ❘}^2=\{\begin{array}{cc}\le 0.5,& \mathrm{belong}\mathrm{to}\mathrm{label}A\\ >0.5,& \mathrm{belong}\mathrm{to}\mathrm{label}B\end{array}& \left(13\right)\end{array}$

[0138]Additionally, in some embodiments, similar to the training process of the quantum neural network, the target data can include N feature data points corresponding to N dimensions, and the feature extraction layer can include K hidden units, where N and K are integers greater than 1.

[0139]On such basis, the aforementioned S420 can specifically include: inputting the N feature data points into each of the K hidden units, and performing feature extraction on the N feature data points using each of the K hidden units to obtain K sub-features, the target data feature including the K sub-features.

[0140]Correspondingly, the above S430 can specifically include: inputting the K sub-features into the unitary matrix layer to obtain K unitary matrices corresponding to the K sub-features.

[0141]Additionally, in some embodiments, the above S440 can specifically include: multiplying the K unitary matrices by the fifth qubit to obtain the sixth qubit.

[0142]It should be noted that the above process is similar to the process for processing the sample data during the above quantum neural network training, which is not repeated herein.

[0151]It should be noted that the above process is similar to the process of obtaining quantum state fidelity during the training of the aforementioned quantum neural network, which is not repeated herein.

[0156]It should be noted that the above process is similar to the process of controlling the swap gate with the first intermediate-state qubit during above quantum neural network training, which is not repeated herein.

[0157]Additionally, in some embodiments, the expression for the unitary matrix layer can also be the aforementioned formula (3). The difference lies in that the input feature θ here can represent the target data feature. On such basis, in some embodiments, values of adjustment parameters α and β can be determined according to the data feature of the target data.

[0158]Combining the various embodiments and implementation methods described above, the quantum neural network structure provided by the present disclosure has the following three advantages.

[0159]A first advantage is that the present disclosure does not involve a quantum state preparation process. There is no need to pre-prepare data into quantum state; instead, data is directly applied to the unitary matrix, and the quantum state of the unit qubit is adjusted using the unitary matrix. The data in the present disclosure is not transformed into quantum state and still remains classical data, and each element in the unitary matrix is also classical data. That is, the classical data does not become quantum state data but becomes the unitary matrix. Each unitary matrix includes four elements, each of which is a real number (a trigonometric function) and has a value range between −1 and 1 (a data range of the trigonometric function). When these unitary matrices multiply qubits, the states of the qubits are changed, i.e., flipped. The classical data is not transformed into the qubits; but rather into tools for flipping qubits.

[0160]A second advantage is that the present disclosure does not require a quantum storage to store the quantum state data. Quantum storage is a major challenge in realizing quantum computing. The quantum system is inherently fragile, and the slightest disturbance from the external world may cause the entire system state to collapse. This characteristic makes quantum state data difficult to store, because it is hard to determine whether the input information has been successfully preserved. Since quantum information cannot be copied or amplified, the quantum storage plays a more important role in quantum information than classical storage does in classical information. In the present disclosure, data is still data, and the unitary matrix also involves conventional data, without involving quantum state data. Therefore, the quantum state data storage is unnecessary. It is important to note that the present disclosure still needs to be executed on a quantum computer, as it involves unit qubit. Thus, quantum storage is still needed to store qubits without storing quantum state data.

[0161]A third advantage is that the present disclosure requires very few qubits (up to 3), resulting in a low cost and preventing the barren plateaus phenomenon. The number of bits in the quantum computer is positively correlated with computational cost, and algorithms requiring more qubits have higher computational costs. For current common quantum neural networks, the number of required qubits is typically P or log P, where P is a feature dimension. Therefore, under a condition that the feature dimension of the sample data is large, for example, 200 features, 200 qubits are typically required. The barren plateaus phenomenon refers to a situation where, when the number of qubits is relatively large (e.g., greater than 10), a framework of traditional quantum neural network easily becomes ineffective for training. The objective function becomes very flat, making gradients difficult to estimate.

[0162]It should be noted that the usage scenarios described in the embodiments of the present disclosure are intended to clarify the technical solutions provided in the embodiments of the present disclosure and do not constitute limitations on the technical solution provided in embodiments of the present disclosure. As can be known to those skilled in the art, with the emergence of new application scenarios, the technical solutions provided by the embodiments of the present disclosure are also applicable to similar technical problems.

[0163]Based on the same inventive concept, the present disclosure provides an apparatus for training a quantum neural network. Detailed description is provided with reference to FIG. 5.

[0164]FIG. 5 is a schematic structural diagram of an apparatus for training a quantum neural network provided by an embodiment of the present disclosure.

[0172]The following provides a detailed description of the apparatus 500 for training the quantum neural network as described above.

[0173]In some embodiments, the sample data includes N feature data points corresponding to N dimensions, and the feature extraction layer includes K hidden units, where N and K are integers greater than 1; the first extraction module 502 is specifically configured to: input the N feature data points into each of the K hidden units, and perform feature extraction on the N feature data points using each of the K hidden units to obtain K sample sub-features, the sample feature including the K sample sub-features; and the first determination module 503 is specifically configured to input the K sample sub-features into the unitary matrix layer to obtain K unitary matrices corresponding to the K sample sub-features.

[0174]In some embodiments, the first adjustment module 504 is specifically configured to multiply the K unitary matrices by the first qubit to obtain the second qubit.

[0185]In some embodiments, the expression for the unitary matrix layer is:

[0187]In some embodiments, a value of each of the adjustment parameters is determined based on a data feature of the sample data.

[0188]Therefore, the structure of the quantum neural network is improved by providing the feature extraction layer, the unitary matrix layer, and the quantum circuit in the quantum neural network. Then, during the training process, the sample feature extracted from the feature extraction layer and corresponding to the sample data is directly input into the unitary matrix layer, the unitary matrix corresponding to the sample feature is prepared to replace the preparation of quantum state data in the traditional method, the first qubit serves as a substitute for the activation function to allow the quantum state of the first qubit to be adjusted to be flipped by the unitary matrix, and after obtaining the second qubit, a similarity comparison between the second qubit and the first qubit is performed by the quantum circuit to calculate the loss function. In this way, in the embodiments of the present disclosure, the quantum neural network can be obtained through training without quantum state preparation, quantum state data storage, or excessive qubits, which can reduce the training difficulty and complexity of the quantum neural network, reduce the computational cost of the quantum neural network, and reduce the probability of occurring of the barren plateaus phenomenon.

[0189]Additionally, the present disclosure also provides an apparatus for classifying data. Detailed description is provided with reference to FIG. 6.

[0190]FIG. 6 is a schematic structural diagram of an apparatus for classifying data provided by an embodiment of the present disclosure.

[0198]The following provides a detailed description of the apparatus 600 for classifying data, and details are presented below.

[0202]In some embodiments, the target data includes N feature data points corresponding to N dimensions, and the feature extraction layer includes K hidden units, where N and K are integers greater than 1; on such basis, the second extraction module 602 is specifically configured to input the N feature data points into each of the K hidden units, and perform feature extraction on the N feature data points using each of the K hidden units to obtain K sub-features, the target data feature including the K sub-features; and the second determination module 603 is specifically configured to input the K sub-features into the unitary matrix layer to obtain K unitary matrices corresponding to the K sub-features.

[0203]In some embodiments, the second adjustment module 604 is specifically configured to multiply the K unitary matrices by the fifth qubit to obtain the sixth qubit.

[0214]In some embodiments, the expression for the unitary matrix layer is:

[0216]In some embodiments, a value of each of the adjustment parameters is determined based on a data feature of the target data.

[0217]Thus, the structure of the quantum neural network is improved by providing the feature extraction layer, the unitary matrix layer, and the quantum circuit in the quantum neural network. Then, during the process for data classification using the quantum neural network, the target data feature extracted from the feature extraction layer and corresponding to the target data to be classified is directly input into the unitary matrix layer, the unitary matrix corresponding to the target data feature is prepared to replace the preparation of quantum state data in the traditional method, the fifth qubit serves as a substitute for the activation function to allow the quantum state of the fifth qubit to be adjusted to be flipped by the unitary matrix, and after obtaining the sixth qubit, a similarity comparison between the sixth qubit and the fifth qubit is performed by the quantum circuit to obtain the quantum state fidelity, and the category to which the target data belongs id determined based on the quantum state fidelity. In this way, in the embodiments of the present disclosure, data classification can be performed using the quantum neural network without quantum state preparation, quantum state data storage, or excessive qubits, which can reduce the training difficulty and complexity of the quantumneural network, and reduce the computational cost of the quantum neural network.

[0218]FIG. 7 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure.

[0219]The electronic device 700 can include a processor 701 and a memory 702 storing computer program instructions.

[0220]Specifically, the processor 701 can include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present disclosure.

[0221]The memory 702 can include a mass storage device for data or instructions. For example, without limitation, the memory 702 can include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or a universal serial bus (USB) drive, or a combination of two or more thereof. Where appropriate, the memory 702 can include removable or non-removable (or fixed) media. Where appropriate, the memory 702 can be located inside or outside an integrated gateway disaster recovery device. In particular embodiments, the memory 702 is non-volatile solid-state storage.

[0222]In particular embodiments, the memory can include a read-only memory (ROM), a random access memory (RAM), a disk storage media device, an optical storage media device, a flash memory device, an electrical, optical, or other physical/tangible memory storage device. Thus, typically, the memory includes one or more tangible (non-transient) computer-readable storage media (e.g., storage devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform operations described in reference to the methods according to one aspect of the present disclosure.

[0223]The processor 701 implements any one of the method for training the quantum neural network and the method for classifying data described in the above embodiments by reading and executing computer program instructions stored in the memory 702.

[0224]In some examples, the electronic device 700 can include a communication interface 703 and a bus 710. As shown in FIG. 7, the processor 701, the memory 702, and the communication interface 703 are connected via the bus 710 to communicate with each other.

[0225]The communication interface 703 is primarily configured to implement the communication between various modules, apparatuses, units, and/or devices in the embodiments of the present disclosure.

[0226]The bus 710 includes hardware, software, or both, which couples the components of an online data traffic metering device together. For example, without limitation, the bus 710 can include an accelerated graphics port (AGP) or other graphics bus, an enhanced industrial standard architecture (EISA) bus, a front side bus (FSB), hyper transport (HT) interconnection, an industrial standard architecture (ISA) bus, unlimited bandwidth interconnect, a low pin count (LPC) bus, a memory bus, a micro channel architecture (MCA) bus, a peripheral component interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a serial advanced technology attachment (SATA) bus, a video electronics standards association local bus (VLB) bus, or other suitable buses, or a combination of two or more thereof. Where appropriate, bus 710 can include one or more buses. Although specific buses are described and illustrated in the embodiments of the present disclosure, the present disclosure takes any suitable bus or interconnect into account.

[0227]Exemplarily, the electronic device 700 can be a mobile phone, a tablet computer, a laptop computer, a handheld computer, an on-vehicle electronic device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), or the like.

[0228]The electronic device 700 can execute the method for training the quantum neural network and/or the method for classifying data described in the embodiments of the present disclosure, thereby implementing the method and apparatus for training the quantum neural network, and/or the method or apparatus for classifying data described in FIGS. 1 to 6.

[0229]Additionally, in conjunction with the method for training the quantum neural network and/or the method for classifying data described in the above embodiments, the present disclosure can provide a computer-readable storage medium for implementation. The computer-readable storage medium stores computer program instructions, and when executed by a processor, the computer program instructions implement any one of the method for training the quantum neural network and/or the method for classifying data described in the above embodiments. Examples of the computer-readable storage media include a non-transitory computer-readable storage medium such as portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disc read-only memory (CD-ROM), optical storage devices, or magnetic storage devices.

[0230]It should be understood that the present disclosure is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, detailed descriptions of known methods have been omitted. In the above embodiments, several specific steps are described and illustrated as examples. However, the method processes of the present disclosure are not limited to the specific steps described and illustrated. Those skilled in the art, upon understanding the gist of the present disclosure, can make various changes, modifications, and additions, or alter the order of the steps.

[0231]The functional blocks illustrated in the above structural block diagram can be implemented as a hardware, a software, a firmware, or a combination thereof. When implemented as hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), an appropriate firmware, a plug-in, a function card, etc. When implemented as software, the elements of the present disclosure are programs or code segments used to perform the required tasks. The programs or code segments can be stored on a machine-readable medium or transmitted on a transmission medium or a communication link via data signals carried by a carrier. “Machine-readable media” can include any medium capable of storing or transmitting information. Examples of machine-readable media include an electronic circuit, a semiconductor memory device, a ROM, a flash memory, an erasable ROM (EROM), a floppy disk, a CD-ROM, an optical disk, a hard disk, a fiber optic media, a radio frequency (RF) link, etc. Code segments can be downloaded via computer networks such as the Internet or intranets.

[0232]It should also be noted that the exemplary embodiments described herein describe certain methods or systems based on a series of steps or apparatuses. However, the present disclosure is not limited to the sequence of steps described above. That is, the steps can be performed in the order described in the embodiments or in an order different from the order described in the embodiments, or multiple steps can be performed simultaneously.

[0233]The various aspects of the present disclosure have been described with reference to flowcharts and/or block diagrams illustrating methods, apparatuses (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each box in the flowcharts and/or block diagrams, and combinations of boxes in the flowcharts and/or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a dedicated computer, or other programmable data processing apparatus to produce a machine, so that the instructions, when executed by the processor of the computer or other programmable data processing apparatus, implement the functions/actions specified in one or more boxes of the flowchart and/or block diagram. Such processor can be, but is not limited to, a general-purpose processor, a dedicated processor, a special application processor, or a field programmable logic device. It is also understood that each block in the block diagram and/or flowchart, as well as combinations of blocks in the block diagram and/or flowchart, can also be implemented by dedicated hardware performing the specified function or action, or by a combination of dedicated hardware and computer instructions.

[0234]The foregoing describes specific embodiments of the present disclosure. Those skilled in the art will readily appreciate that, for the sake of description and brevity, the specific operational processes of the systems, modules, and units described above can be understood with reference to the corresponding processes in the preceding method embodiments and are not repeated herein. It should be understood that the protection scope of the present disclosure is not limited to the above. Those skilled in the art can readily conceive various equivalent modifications or replacements within the scope of the disclosed technology, and such modifications or replacements should fall within the protection scope of the present disclosure.